[garth-wells]

Replaces the mypy -p ffcx CI step with pyrefly check, and fixes the type errors that surfaced.

Two commits: the CI and configuration swap, then the source changes it required.

Configuration

[tool.mypy] becomes [tool.pyrefly], mapping each option across:

mypy	pyrefly
check_untyped_defs	check-unannotated-defs
disallow_untyped_defs	implicit-any-parameter, unannotated-return
warn_unused_ignores	unused-ignore
per-module overrides	[[sub-config]] on ffcx/ir/**, ffcx/codegeneration/**

The exempted modules use replace-imports-with-any rather than ignore-missing-imports. The latter applies only when a module cannot be found, so the result would depend on the platform and on how the dependency was installed: pygraphviz is installed on Linux only, and UFL resolves when pip-installed from git but not from an editable checkout. Checking against UFL's own types reports 61 errors where treating it as Any reports 45, so it is replaced with Any, as the mypy configuration effectively did. basix and numba are now checked for real, which costs one further error and caught a wrong return annotation.

types-setuptools is dropped: it was a stub package for mypy, nothing in ffcx imports setuptools, and pyrefly carries its own typeshed.

unbound-name is off. mypy's equivalent is opt-in and was not enabled, so turning it on would newly fail CI. It reports four sites; three are guarded twice on the same condition, but the t in elementtables.py looks genuinely reachable — an unhandled 3D cell type falls through with it unset.

Fixes

Six were real:

• access.py used False as the not-found sentinel for a dict of handlers, so the value type was Callable | bool and the truthy branch was not callable. Now None, as definitions.py already does — this also drops a # type: ignore.
• interface.py declared form_generator returning tuple[str]. The C backend returns (declaration, implementation); only numba returns a 1-tuple. The sibling protocols already use tuple[str, ...].
• numba/utils.py declared get_void_pointer returning a RawPointer. numba.types.voidptr(arr) builds a Signature, as its own docstring says.
• elementtables.py returned a bare dict of the table and its offsets, so reading the offsets back gave ndarray | int. It now has a TypedDict.
• optimizer.py's _drop_unreferenced returns whatever shape it is given; overloads say so.
• representation.py passed tensor_shape, a list, where a tuple is declared and only its length is read.

Three are suppressed, all the same thing: **-unpacking a dict into a NamedTuple, which the checker can follow neither for the field set nor for their types.

Why pyrefly is floored at 1.4

Worth knowing before the next upgrade. pyrefly 1.3 additionally reports singledispatchmethod making __call__ a descriptor where the protocol declares a plain method, and an isinstance that cannot narrow; 1.4 reports neither. Because unused-ignore is on — mypy's warn_unused_ignores, which this project had — suppressing them fails on 1.4 and not suppressing them fails on 1.3. There is no source that satisfies both, so the floor picks the version whose answer the source matches. Verified: clean on 1.4.0, 3 errors on 1.3.2.

This cuts both ways: a later release that fixes one of the three suppressions above will fail CI the same way until that suppression is removed. That is unused-ignore doing its job, but it does couple the repo to the checker version.

One bug left alone

ffcx/ir/integral.py compares p["part"], the option string "full"/"diagonal", against TensorPart, a plain Enum over ints. That comparison is always False, so its branch never runs. Comparing TensorPart.from_str(p["part"]) instead sets rank = 1 and fails all 18 part="diagonal" tests, so the live behaviour is load-bearing and the intent needs deciding rather than the comparison simply being repaired. The call site carries a FIXME; filed as #929.

Verification

pyrefly check reports 0 errors on 1.4.0, both with UFL resolvable and not, and exits 1 on a deliberately introduced error so the CI step fails when it should. Full suite 474 passed, demos 232 passed, ruff and ruff format clean.

Note this touches elementtables.py, which #903 restructures — they will conflict.

🤖 Generated with Claude Code

[schnellerhase]

Why replace and not add as an additional check? The two catch different things.

[garth-wells]

Why replace and not add as an additional check? The two catch different things.

I don't see that two tools (or 3 or 4, or ..) brings added value in this case. We use pyrefly only for DOLFINx and Basix. This provides consistency across the projects.

[schnellerhase]

Mypy is more mature, if only one tool is to be picked, I think that's the one. We have been running in dolfiny ty, mypy and pyrefly, they do not collide and do catch different things in certain cases. More coverage is always preferable.